NVIDIA/nvidia-kaggle

NVIDIA Kaggle Plugin gives agents end-to-end Kaggle competition workflows through a single skill, nvidia-kaggle-skill. It can gather competition context, study public writeups and notebooks, reproduce kernels locally, submit to competitions, and manage Ka

NVIDIA Kaggle Plugin – Agent Skill for End‑to‑End Kaggle Workflows

What it is – A Python‑based Agent Skill that lets large‑language‑model agents (e.g., Claude Code, Codex) talk to Kaggle. With a single natural‑language command (nvidia‑kaggle‑skill) the agent can:

  • Pull competition overviews, rules, and dataset metadata.
  • Search and summarize leaderboard write‑ups, discussion threads, and public notebooks.
  • Download a notebook (kernel) together with its data into a local workspace ready to run.
  • Submit a notebook to a competition and poll for the result.
  • Create or update a Kaggle dataset from a local folder.

The skill parses whatever you give it – a competition slug, a Kaggle URL, a kernel reference, a write‑up link, or a local path – and automatically selects the appropriate sub‑script.

Key capabilities (as shown in the README)

Capability What the skill does
Competition context Returns overview, rules, evaluation metric, timeline, and dataset description.
Solution write‑ups Finds top‑ranked leaderboard write‑ups and produces a concise summary of the winning approaches.
Discussions Indexes Kaggle discussion threads, lets the agent search them, and reads selected posts.
Kernels Indexes public kernels, queries them, and extracts the most relevant notebooks.
Kernel reproduction Downloads a public notebook plus all required inputs and arranges them in a local directory ready for execution.
Submission Pushes a notebook‑based submission to Kaggle and polls the competition page for the final score.
Dataset upload Creates a new Kaggle dataset or updates an existing one from a local folder.

Quick start (installing the skill)

  1. Prerequisites
    • Python 3.10+.
    • An agent runtime that supports plugins/Agent Skills (e.g., Codex, Claude Code).
    • A Kaggle account with a KAGGLE_API_TOKEN environment variable.
  2. Marketplace install (for supported runtimes)
    # Codex
codex plugin marketplace add https://github.com/NVIDIA/nvidia-kaggle.git
# Claude Code
claude plugin marketplace add https://github.com/NVIDIA/nvidia-kaggle.git
claude plugin install nvidia-kaggle@nvidia-kaggle --scope user
  3. Local install (any runtime that can read a skill directory)
    cp -R skills/nvidia-kaggle-skill <your‑skills‑directory>/
    The directory must contain SKILL.md, the markdown workflow files, and the scripts/ folder.

Example usage patterns (from the README)

  • Summarize top solutions
    /nvidia-kaggle:nvidia-kaggle-skill Get the top 3 solution writeups from the AI Mathematical Olympiad – Progress Prize 2 competition and summarize the key strategies.
  • Fetch competition overview & dataset description
    /nvidia-kaggle:nvidia-kaggle-skill Fetch the competition overview and dataset description for the ARC Prize 2025 competition.
  • Research public kernels
    /nvidia-kaggle:nvidia-kaggle-skill Research the top public kernels for the Home Credit Default Risk competition and summarize the modeling approaches used.
  • Set up a notebook locally
    /nvidia-kaggle:nvidia-kaggle-skill Download and set up this notebook locally so I can run it: https://www.kaggle.com/code/cdeotte/titanic-wcg-xgboost-0-84688

These commands illustrate the natural‑language interface: the agent decides which internal script to run based on the request.

Development & testing

  • Run the test suite with uv run pytest.
  • Integration tests (which hit the real Kaggle API) are gated behind --run-integration and require a valid token.
  • Before releasing, sync dependencies (uv sync), run the full test suite, and validate the plugin with claude plugin validate ..

License

  • MIT (see LICENSE).

Bottom line – The NVIDIA Kaggle Plugin turns a conversational AI agent into a fully‑featured Kaggle assistant, handling everything from data discovery to model submission without the user writing any API code themselves.

Related

  • Project
    Kaggle/kaggle-cli

    The official command-line interface for Kaggle, allowing users to manage datasets, competitions, models, and notebooks directly from their terminal.

  • Project
    Kaggle/kagglehub

    A Python library for programmatically downloading, uploading, and loading Kaggle datasets, models, and competition resources into data science frameworks.

  • Project
    dongshuyan/compass-skills

    COMPASS Skills is an open‑source collection of eight local SKILL.md packages that give AI agents concrete tools for clarifying ambiguous requests, maintaining a repo‑wide task DAG, creating hand‑off prompts, storing a user collaboration profile, engineering and upgrading reusable skills, humanising academic prose, and preparing structured interview reports. Installable via `npx skills add …` for Claude Code, Codex, etc., the skills run entirely locally (Python standard library only) and require explicit user approval for any high‑risk operation. The project is in alpha, MIT‑licensed, and includes unit tests and smoke‑test validation.

  • Project
    Neeeophytee/finding-unknowns-skills

    A collection of 11 installable “skills” (SKILL.md files) for Claude Code, OpenAI Codex, Kimi, Hermes, Cursor, etc. They help developers surface unknowns, plan implementations, and keep agent context clean. Install with a single `npx skills add …` command or via each agent’s plugin system.

  • Project
    kangarooking/cangjie-skill

    A framework for distilling methodologies from books, videos, and podcasts into structured, executable AI skills that agents can invoke for real-world problem solving.